Using reading history and AI to find better book recommendations.
User demographics
Exploratory interviews revealed Goodreads as the most cited social media platform for online book related activity, confirming it as a good platform to build for. Based on Goodreads traffic reports and largely similar survey respondent populations, the following parameters were made to create representative groups for subsequent randomized user testing panels of five participants:
GENDER
3 women
2 men
AGE
Age 18-24: 1 participant
Age 25-35: 2 participants
Age 35-44: 1 participant
Age 45-54: 1 participant
Books per year read
5-11 books: 1 participant
12-24 books: 3 participants
25-49 books: 1 participant
Country
USA
AI usage
Occasional to frequent
Reading content (genres varied)
~90% fiction
~10% nonfiction
Problem
A majority of readers have trouble finding books they like, using diverse and often time-consuming methods, often deferring to social media, and AI chatbots for more specific recommendations.
Constraints
Exploratory research confirmed Goodreads as an appropriate platform for build. Time and budget were limited.
Exploratory research
Screening and participant selection
Screener survey was built to assess demographics, AI use, reading volume and genres, and difficulty finding books.
Eight participants were selected to capture broad array of reading volumes and styles, with some curve toward Goodreads user behaviors (shown in survey responses as a significantly used platform) and AI familiarity (as most respondents were frequent to occasional users of AI). Survey respondent population was also largely mirroring Goodreads demographics and behavior. Most survey respondents had difficulty finding books, and participants from that pool were selected for interview.
Reader behavior findings
From the pool of eight respondents, a majority find books through Goodreads and friends (3 of 8, each), followed closely by ChatGPT, Gemini, bookstores, Storygraph, and TikTok (2 of 8, each).
Everyone was open to sharing with AI, while 2 of 8 actively share their reading history with AI or digital tools to get recommendations. 3 of 8 mentioned not wanting personally identifying information shared (including mood), but 2 of 8 said they are comfortable sharing around reading habits specifically. 1 of 8 wanted to know data and confidentiality policies.
5 of 8 said recommendations were hit-or-miss, while individuals had particular tools or methods they found useful (social media, digital services, bookstores). Social media was trusted the most for recommendations (3 of 8), with Goodreads mentioned 2x as any other. Otherwise Wattpad, AI, bookstores, and Amazon reviews were individually mentioned.
6 of 8 use ChatGPT or Gemini when looking for something specific, with 3 of 8 saying emotion of books was important. Methods for searching were otherwise mixed.
Ideation
Taxonomy-based user flow
Devised simple chatbot user flow with options to define by the common search criteria, also surfacing these category details in the results to allow user to deeply search
Created clickable with Claude from wireframes
Pros: Prompts user with common search criteria, allows user to use categoric searching for discovery
Cons: Complicated interface, requires extra attention of user, provides information user never asks for, perhaps overcomplicating
Basic user flow for AI chatbot
This operated mainly on a refined object-action pair within a simplified chatbot archetype to align with constraints, existing user behavior and familiarity
Pros
• Users already using this archetype to search for books
• Intuitive, simplified
• Does not categorical thinking on user
Cons
• Requires user input for results
AI chatbot
Understanding basic chatbot user flow and elaborating as feature for Goodreads app.
Pros:
• Familiar chatbot archetype, established usage pattern
• Goodreads conventions for exploring recommendations, no new mental models
• AI privacy and logic details made obvious
• “Use my reading history” made obvious
• Clear headline to signal function and call to action
• Input field provides pretext for user to elaborate on
Cons:
• Requires user thinking and input for recommendations
“Automatic” variations
One of the usability testing participants of the clickable prototype found thinking and typing to get recommendations to be too complicated. However they did like the recommendations based on reading history, and ability to refine recommendations with a chatbot. This pointed to developing several options shown below.
1. Automatic list with chatbot to refine
Condensing the original chatbot flow to automatically generate recommendations list based on user's reading history, with chatbot at bottom to allow user to adjust the recommendations further.
Pros:
• Does not require input from user for recommendations
• Could be surfaced in AI function button
Cons:
• More programmatic of user, leading
2. "Based on your reading history" carousel
Adding this carousel to the Goodreads "Discover" mode, where users are familiar with seeing recommendations carousels based on various topics.
Pros:
• Could be evergreen to support search
• Does not require input from user for recommendations
Cons:
• Limited functionality
• Easier to miss
• More programmatic of user, leading
Usability testing
Testing of AI chatbot clickable prototype was conducted in two rounds, revealing areas for improvement and possible need for automatically populated recommendations.
Round 1
After 2 interviews, found users fumbling during text entry and advancement on clickable prototype.
• Stop testing and quick fix: add "send" icon to main input field. (Originally, field was automatically populated with dummy text with no popup keyboard represented, breaking mobile conventions. Adding a "send" icon made necessary action obvious, whereas in a more functional setup the "return" key would be pressed on the popup keyboard to proceed.)
Round 2
All users reached goal without error. Synthesis revealed user sentiment and next steps:
• Refinements to clickable prototype
• Alternate directions for automatic recommendations
